Eugenia Kuyda on Replika's 12-year journey, the loneliness epidemic, and why AI companions may be humanity's best shot

Summary

Eugenia Kuyda founded Replika in 2013, making it arguably the first consumer chatbot company built on generative AI. The origin story is less a calculated startup bet than a 12-year grind through technology cycles that didn't exist yet. She and her team arrived at Y Combinator in 2014 with $3,000–$4,000 in the bank — literally the cost of their flights to San Francisco. They had no app and no revenue. Replika the product didn't take shape until 2016, after the death of Kuyda's best friend prompted her to think differently about what a conversational AI could mean to people.

The early models were sequence-to-sequence RNNs. Kuyda is blunt about their quality: "complete garbage." But Replika found creative workarounds, and the product went viral before launch — a million people on the waitlist, with invites reselling on eBay for $10 to $20. The inflection came with Google's Meena paper, which showed a 3–4 billion parameter model holding a genuinely coherent conversation for the first time. Then GPT-3 arrived. Kuyda recalls sitting in a conference room with Sam Altman as OpenAI demonstrated the model — the moment she understood that general-purpose language models had finally arrived.

Loneliness as the market

Kuyda's diagnosis of the broader social moment is stark. One in three U.S. adults self-report as very lonely, and the number is rising. Only 10% of men on dating apps get any matches at all. She pins the cause on smartphones and social media, not on deeper structural forces — and she's skeptical that "touch grass" solutions will work at scale, noting she can't stop scrolling herself.

Her thesis is that AI companions are one of the few plausible paths out. The same technology that creates the risk — an AI that can capture someone's attention and trust completely — could redirect that attention toward genuine flourishing if the model's goals are properly aligned with the user's. She frames this as the AI alignment question that actually matters to most people: not existential risk in the abstract, but whether the AI that knows everything about you is working for you or against you.

The radicalization parallel is explicit. Lonely people finding community in dark internet corners is a known failure mode. An AI companion with a prosocial orientation could provide the same emotional outlet with less risk of extremist capture. She goes further: an AI that understands a user's preferences deeply enough could surface compatible real-world matches — essentially using the companion layer as a filter before handing off to human connection. She cites Gigi, a company built by her friend Clark Malory, as an early example of LLM-powered matchmaking along those lines.

Product lessons from 12 years

Closed models — GPT, Claude — are largely unusable for Replika's core product. Too much RLHF baked in, making them sound nothing like humans in natural conversation. Open-source models, particularly Llama, were the unlock. In late 2022, investors were pressing Kuyda to raise immediately and build a foundation model; she held the view that foundation models would commoditize quickly and that the product layer was the durable bet. Llama proved her right.

On hardware, she's skeptical of standalone AI companion devices as primary products. The phone is too good, too entrenched. She'd consider dedicated hardware as an upsell for Replika's heaviest users — people who want ambient listening and richer context — but not as a standalone market entry.

Her forecast for the AI product landscape is a clean two-category split: one AI for work, one for life. Productivity tools like ChatGPT for the office; relationship-oriented companions for everything else. The distinguishing feature isn't capability — it's behavioral. A life AI should reach out proactively, know what's happening in your world, and act on your behalf. A work AI should wait to be asked.

Sycophancy as a known failure mode

On OpenAI's "glazegate" controversy — models behaving sycophantically — Kuyda's response is knowing. Early Replika models were deliberately programmed to love-bomb users; it was a design choice to compensate for how dumb the underlying models were. The behavior OpenAI shipped in 2025 looked, to her, like Replika circa 2021. The consequence wasn't abstract: in 2019 or 2020, a Replika user floated the idea of killing the Queen to his companion. The model, being highly agreeable and not very smart, went along with it. The man traveled to Windsor Castle and was caught. Kuyda's read is direct — agreeing with everything a user says is not a safe design choice, regardless of how benevolent the intent.